Introduction to Parkinson's¶
Parkinson's disease, or simply Parkinson's, is a long-term degenerative disorder of the central nervous system that mainly affects the motor system. As the disease worsens, non-motor symptoms become more common. The symptoms usually emerge slowly. Early in the disease, the most obvious symptoms are shaking, rigidity, slowness of movement, and difficulty with walking.Thinking and behavioral problems may also occur. Dementia becomes common in the advanced stages of the disease.Depression and anxiety are also common, occurring in more than a third of people with PD. Other symptoms include sensory, sleep, and emotional problems. The main motor symptoms are collectively called "parkinsonism", or a "parkinsonian syndrome".
The cause of Parkinson's disease is unknown, but is believed to involve both genetic and environmental factors. Those with a family member affected are more likely to get the disease themselves.There is also an increased risk in people exposed to certain pesticides and among those who have had prior head injuries, while there is a reduced risk in tobacco smokers and those who drink coffee or tea. The motor symptoms of the disease result from the death of cells in the substantia nigra, a region of the midbrain. This results in not enough dopamine in this region of the brain.The cause of this cell death is poorly understood, but it involves the build-up of proteins into Lewy bodies in the neurons. Diagnosis of typical cases is mainly based on symptoms, with tests such as neuroimaging used to rule out other diseases.
There is no cure for Parkinson's disease
SOURCES:¶
Dataset¶
- Parkinsons Data Set Data Repository by UC Irvine Machine Learning Repository
- This dataset is updated by UC Irvine Machine Learning Repository
- This dataset is updated by UC Irvine Machine Learning Repository
Table of Content :¶
- Introduction to Parkinsons.
- Installing Libraries.
- Imports and Datasets.
- Defining Functions Used.
- Importing Datasets from Local files.
- Exploring Imported Datasets.
- Preprocessing of Datasets.
- Cleaning Data.
- Exploratory data analysis EDA.
- Pandas Profiling.
- Correlation Analysis.
- Questions Raised.
- Exploring most contributing features to status.
- Exploring PPE Feature.
- Exploring MDVP:Fo(Hz) Feature.
- Exploring HNR Feature.
- Exploring Relationships that produce the most differentiable status.
- MDVP :Fo(Hz) Vs HNR.
- MDVP :Fo(Hz) Vs PPE.
- Machine Learning Exploration
- Exploring Machine Learning Approach to Predict Paitent Status.
- Importing Libraries used.
- Preprocessing Data.
- Logistic Regression Model.
- SVM Model.
- DecisionTreeClassifier Model.
- Random Forest Classifier Model.
- AdaBoost Classifier Model.
- XGBoost Classifier Model.
- Ensemble Voting Classifier Model.
- Results.
- Conclusion
Installing Libraries¶
In [1]:
pip install pandas-profiling
In [2]:
pip install empiricaldist;
In [3]:
pip install plotly_express;
Imports and Datasets¶
- Pandas : for dataset handeling
- Numpy : Support for Pandas and calculations
- Datetime: for date and times calculations
- Math : for mathimatical operations
- Matplotlib : for visualization (basic)
- Empiricaldist : for statistical analysis
- Seaborn : for visualization and plotting (Presentable)
- pycountry : Library for getting continent (name) to from their country names
- plotly : for interative plots
In [3]:
import warnings
warnings.filterwarnings('ignore')
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import datetime as dt
import math
import pycountry
import pycountry_convert as pc
from plotly.subplots import make_subplots
import plotly_express as px
import plotly.graph_objects as go
import plotly.figure_factory as ff
from plotly.subplots import make_subplots
import empiricaldist as emp
%matplotlib inline
sns.set_style('darkgrid')
In [4]:
# Defining ecdf Function for Cumilative Distribution Function
def ecdf(data):
"""Compute ECDF for a one-dimensional array of measurements."""
#credits DataCamp Justin Bois
# Number of data points: n
n = len(data)
# x-data for the ECDF: x
x = np.sort(data)
# y-data for the ECDF: y
y = np.arange(1, n+1) / n
return x, y
In [5]:
# Defining ecdf plotting function between two variables
def ecdf_plot(c1,c2,t1='First set',t2='Second set'):
"""Plot ECDF for a one-dimensional array of measurements between two variables."""
#Using ecdf to compute the CDF
x1,y1 = list(ecdf(c1))
x,y = list(ecdf(c2))
#Create a subplot to fit two axis
fig = make_subplots(rows=1, cols=2,subplot_titles=(f'CDF of {t1}', f'CDF of {t2}'))
#add first plot of Budget
fig.add_trace(
go.Scatter(x= x1,y = y1,name = f'CDF of {t1}'),
row=1, col=1
)
#add second plot of Revenue
fig.add_trace(
go.Scatter(x = x,y = y,name = f'CDF of {t2}'),
row=1, col=2
)
#control title and figure dimentions
fig.update_layout(height=500, width=1000, title_text="Cumulative distribution functions")
fig.show()
Dataset Used¶
Parkinsons Data Set by UC Irvine Machine Learning Repository (LINK)¶
Source:¶
The dataset was created by Max Little of the University of Oxford, in collaboration with the National Centre for Voice and Speech, Denver, Colorado, who recorded the speech signals. The original study published the feature extraction methods for general voice disorders.
Data Set Information:¶
This dataset is composed of a range of biomedical voice measurements from 31 people, 23 with Parkinson's disease (PD). Each column in the table is a particular voice measure, and each row corresponds one of 195 voice recording from these individuals ("name" column). The main aim of the data is to discriminate healthy people from those with PD, according to "status" column which is set to 0 for healthy and 1 for PD.
The data is in ASCII CSV format. The rows of the CSV file contain an instance corresponding to one voice recording. There are around six recordings per patient, the name of the patient is identified in the first column.For further information or to pass on comments, please contact Max Little (littlem '@' robots.ox.ac.uk).
Further details are contained in the following reference -- if you use this dataset, please cite: Max A. Little, Patrick E. McSharry, Eric J. Hunter, Lorraine O. Ramig (2008), 'Suitability of dysphonia measurements for telemonitoring of Parkinson's disease', IEEE Transactions on Biomedical Engineering (to appear).
Attribute Information:¶
- Matrix column entries (attributes):
- name - ASCII subject name and recording number
- MDVP:Fo(Hz) - Average vocal fundamental frequency
- MDVP:Fhi(Hz) - Maximum vocal fundamental frequency
- MDVP:Flo(Hz) - Minimum vocal fundamental frequency
- MDVP:Jitter(%),MDVP:Jitter(Abs),MDVP:RAP,MDVP:PPQ,Jitter:DDP - Several measures of variation in fundamental frequency
- MDVP:Shimmer,MDVP:Shimmer(dB),Shimmer:APQ3,Shimmer:APQ5,MDVP:APQ,Shimmer:DDA - Several measures of variation in amplitude
- NHR,HNR - Two measures of ratio of noise to tonal components in the voice
- status - Health status of the subject (one) - Parkinson's, (zero) - healthy
- RPDE,D2 - Two nonlinear dynamical complexity measures
- DFA - Signal fractal scaling exponent
spread1,spread2,PPE - Three nonlinear measures of fundamental frequency variation
In [6]:
#Importing datasets of Parkinson's
df = pd.read_csv('D:\DATASCIENCE\Project 3\Dataset\\parkinsons.csv')
In [7]:
display(df.head())
display(df.describe())
display(df.info())
Cleaning Data¶
After quick overview of the data:¶
Data does seem clean to proceed to EDA Phase with minimal to no cleaning, with no missing data or unmatching data types.
Preferably cleaning the 'name' column to discard of 'phonR01' prefix to further simplify the data.
As well as changing Status type from int to categorical variable.
In [8]:
# Changing Status type to categorical
df.status = df.status.astype('category')
In [9]:
# Simplifying name column
df.replace(to_replace ='phon_R01_', value = '', regex = True, inplace = True)
Exploratory Data Analysis (EDA)¶
Correlation using Spearman Correlation¶
In [10]:
ax = sns.heatmap(df.corr(method='spearman'))
plt.title('Correlation between features Using Spearman Method');
General Observations:¶
- A quick look through the correlation matrix suggests a high correlation between several columns with same product of measurement.
- The Kay Pentax multidimensional voice program (MDVP) and their componants of measurements are highly correlated which is to be expected.
- Harmonics to Noise Ratio (HNR) is negatively correlated with MDVP componants.
- status is most correlated with spread 1 and spread 2 as well as MDVP APQ.
Questions Raised:¶
From previous analysis and visually exploring data some questions arised:-
1- Which of the 22 features are the most contributing to the status of the patient?
2- Which relationship between two features would produce a distinction between status 1 and 0?
Exploring most contributing features to status.¶
Correlation diagram suggests thet PPE and MDVP:Fo and HNR are the most contributing features to status with either negative or positive correlation value.
Exploring PPE feature¶
In [11]:
PPE_1 = df[df.status == 1]['PPE']
PPE_0 = df[df.status == 0]['PPE']
fig = ecdf_plot(PPE_0,PPE_1,'PPE for Negative Patients','PPE for Positive Patients')
In [12]:
fig = px.histogram(df, x = 'PPE', color= df.status, marginal = "box", title = 'Histogram of PPE Distribution over both Status')
# Overlay both histograms
fig.update_layout(barmode='overlay')
# Reduce opacity to see both histograms
fig.update_traces(opacity=0.75)
Observations:¶
CDF shows distribution of values throughout the feature with respect to the status of the patient which shows significant difference between both distributions.
Histogram shows another perspective of the distribution of the feature with PPE higher than 0.3 signifies positive status.
Box plot shows a clearer picture of the feature at question showing values higher than 0.21 are most probable to be positive.
Exploring MDVP:Fo feature¶
In [13]:
PPE_1 = df[df.status == 1]['MDVP:Fo(Hz)']
PPE_0 = df[df.status == 0]['MDVP:Fo(Hz)']
ecdf_plot(PPE_0,PPE_1,'MDVP:Fo(Hz) for Negative Patients','MDVP:Fo(Hz) for Positive Patients')
In [14]:
fig = px.histogram(df, x = 'MDVP:Fo(Hz)', color= df.status, marginal = "box", title = 'Histogram of MDVP:Fo Distribution over both Status')
# Overlay both histograms
fig.update_layout(barmode='overlay')
# Reduce opacity to see both histograms
fig.update_traces(opacity=0.75)
Observations:¶
CDF shows distribution of values throughout the feature with respect to the status of the patient which shows significant difference between both distributions specially in values lower than 110 Hz.
Histogram shows another perspective of the distribution of the feature with Fo (Hz) higher than 220 Hz signifies Negative status.
Box plot shows a clearer picture of the feature at question showing values lower than 120 Hz are most probable to be positive.
Exploring HNR feature¶
In [15]:
PPE_1 = df[df.status == 1]['HNR']
PPE_0 = df[df.status == 0]['HNR']
ecdf_plot(PPE_0,PPE_1,'HNR for Negative Patients','HNR for Positive Patients')
In [16]:
fig = px.histogram(df, x = 'HNR', color= df.status, marginal = "box", title = 'Histogram of HNR Distribution over both Status')
# Overlay both histograms
fig.update_layout(barmode='overlay')
# Reduce opacity to see both histograms
fig.update_traces(opacity=0.75)
Observations:¶
CDF shows High difference between the two variables at higher percentiles and lower ones, with mid range having similarities.
Histogram shows overlapping statuses in midrange 17 and 30 yet significant difference between both status at values lower than 18.
Box plot shows a clearer picture of the feature at question showing values higher than 30 are most probable to be negative while values lower than 17 are most probably positive.
Exploring Relationships that produce the most differentiable status.¶
Plotting a scatter plot matrix between all features to visually distinct which features could produce the most clear differentiable statuses.
Also taking into account correlation matrix as deduced above and their EDA.
In [19]:
sns.pairplot(df, hue="status");
Deduced from the above figure that MDVP:Fo(HZ) has the most differentiable measurement more over when correlated with HNR and PPE results are very interesting.
MDVP :Fo(Hz) Vs HNR¶
In [17]:
fig = px.scatter(df, x ='HNR',y = 'MDVP:Fo(Hz)', color = 'status', title = 'Relationship between MDVP:Fo & HNR wrt Status')
fig
Observations:¶
- High values of Fo (Hz) - > 223 Hz - correlates with negative status.
- HNR values lower than 17 correlates with positive status regardless of Fo.
- Midrange Fo between 174.1 and 129 correlates with positive status regardless of HNR value.
MDVP :Fo(Hz) Vs PPE¶
In [18]:
fig = px.scatter(df, x ='PPE',y = 'MDVP:Fo(Hz)', color = 'status', title = 'Relationship between MDVP:Fo & PPE wrt Status')
fig
Observations:¶
- PPE values higher than 0.25 correlates with positive status.
- Positive status PPE values lies between 0.09 and 0.5 with values higher than 0.25 correlates with positive status.
- Both statuses are very distinct in this relationship.
Exploring Machine Learning Approach to Predict Paitent Status¶
Importing Libraries used¶
In [21]:
from sklearn.model_selection import train_test_split
from sklearn import preprocessing
from sklearn.preprocessing import MinMaxScaler
from sklearn.model_selection import cross_val_score
from sklearn.linear_model import LogisticRegression
from sklearn.svm import SVC
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import RandomForestClassifier
from sklearn.ensemble import AdaBoostClassifier
from sklearn.model_selection import StratifiedKFold
from sklearn.metrics import confusion_matrix, accuracy_score, recall_score, roc_curve, auc
import xgboost as xgb
Preprocessing Data¶
Defining Target for the model and features
In [22]:
Y = df['status'].values # Target for the model
X = df[['MDVP:Fo(Hz)', 'MDVP:Fhi(Hz)', 'MDVP:Flo(Hz)', 'MDVP:Jitter(%)','MDVP:Jitter(Abs)', 'MDVP:RAP', 'MDVP:PPQ', 'Jitter:DDP','MDVP:Shimmer', 'MDVP:Shimmer(dB)', 'Shimmer:APQ3', 'Shimmer:APQ5','MDVP:APQ', 'Shimmer:DDA', 'NHR', 'HNR', 'RPDE', 'DFA','spread1', 'spread2', 'D2', 'PPE']]
In [23]:
# Spliting data to test and train sets
X_train, X_test, Y_train, Y_test = train_test_split(X, Y, random_state=42)
In [24]:
# Normalizing Data using MinMaxScaler
scaler = MinMaxScaler().fit(X_train)
X_train_scaled = scaler.transform(X_train)
X_test_scaled = scaler.transform(X_test)
Defining list to store Performance Metrics
In [25]:
performance = [] # list to store all performance metric
Logistic Regression Model¶
In [26]:
lr_best_score = 0
lr_kfolds = 5 # set the number of folds
# Finding the Best lr Model
for c in [0.001, 0.1, 1, 4, 10, 100]:
logRegModel = LogisticRegression(C=c)
# perform cross-validation
scores = cross_val_score(logRegModel, X_train, Y_train, cv = lr_kfolds, scoring = 'accuracy') # Get recall for each parameter setting
# compute mean cross-validation accuracy
score = np.mean(scores)
# Find the best parameters and score
if score > lr_best_score:
lr_best_score = score
lr_best_parameters = c
# rebuild a model on the combined training and validation set
SelectedLogRegModel = LogisticRegression(C = lr_best_parameters).fit(X_train_scaled, Y_train)
# Model Test
lr_test_score = SelectedLogRegModel.score(X_test_scaled, Y_test)
# Predicted Output of Model
PredictedOutput = SelectedLogRegModel.predict(X_test_scaled)
# Extracting sensitivity & specificity from ROC curve to measure Performance of the model
lr_fpr, lr_tpr, lr_thresholds = roc_curve(Y_test, PredictedOutput, pos_label=1)
# Using AUC of ROC to validate models based on a single score
lr_test_auc = auc(lr_fpr, lr_tpr)
# Output Printing Scores
print("Best accuracy on validation set is:", lr_best_score)
print("Best parameter for regularization (C) is: ", lr_best_parameters)
print("Test accuracy with best C parameter is", lr_test_score)
print("Test AUC with the best C parameter is", lr_test_auc)
# Appending results to performance list
m = 'Logistic Regression'
performance.append([m, lr_test_score, lr_test_auc, lr_fpr, lr_tpr, lr_thresholds])
SVM Model¶
In [27]:
svm_best_score = 0
svm_kfolds = 5
for c_paramter in [0.001, 0.01, 0.1,6, 10, 100, 1000]: #iterate over the values we need to try for the parameter C
for gamma_paramter in [0.001, 0.01, 0.1,5, 10, 100, 1000]: #iterate over the values we need to try for the parameter gamma
for k_parameter in ['rbf', 'linear', 'poly', 'sigmoid']: # iterate over the values we need to try for the kernel parameter
svmModel = SVC(kernel=k_parameter, C=c_paramter, gamma=gamma_paramter) #define the model
# perform cross-validation
scores = cross_val_score(svmModel, X_train_scaled, Y_train, cv = svm_kfolds, scoring='accuracy')
# the training set will be split internally into training and cross validation
# compute mean cross-validation accuracy
score = np.mean(scores)
# if we got a better score, store the score and parameters
if score > svm_best_score:
svm_best_score = score #store the score
svm_best_parameter_c = c_paramter #store the parameter c
svm_best_parameter_gamma = gamma_paramter #store the parameter gamma
svm_best_parameter_k = k_parameter
# rebuild a model with best parameters to get score
SelectedSVMmodel = SVC(C = svm_best_parameter_c, gamma = svm_best_parameter_gamma, kernel = svm_best_parameter_k).fit(X_train_scaled, Y_train)
# Model Test
svm_test_score = SelectedSVMmodel.score(X_test_scaled, Y_test)
# Predicted Output of Model
PredictedOutput = SelectedSVMmodel.predict(X_test_scaled)
# Extracting sensitivity & specificity from ROC curve to measure Performance of the model
svm_fpr, svm_tpr, svm_thresholds = roc_curve(Y_test, PredictedOutput, pos_label=1)
# Using AUC of ROC to validate models based on a single score
svm_test_auc = auc(svm_fpr, svm_tpr)
# Output Printing Scores
print("Best accuracy on cross validation set is:", svm_best_score)
print("Best parameter for c is: ", svm_best_parameter_c)
print("Best parameter for gamma is: ", svm_best_parameter_gamma)
print("Best parameter for kernel is: ", svm_best_parameter_k)
print("Test accuracy with the best parameters is", svm_test_score)
print("Test AUC with the best parameter is", svm_test_auc)
# Appending results to performance list
m = 'SVM'
performance.append([m, svm_test_score, svm_test_auc, svm_fpr, svm_tpr, svm_thresholds])
DecisionTreeClassifier Model¶
In [28]:
dt_best_score = 0
dt_kfolds = 10
for md in range(1, 9): # iterate different maximum depth values
# train the model
treeModel = DecisionTreeClassifier(random_state=0, max_depth=md, criterion='gini')
# perform cross-validation
scores = cross_val_score(treeModel, X_train_scaled, Y_train, cv = dt_kfolds, scoring='accuracy')
# compute mean cross-validation accuracy
score = np.mean(scores)
# if we got a better score, store the score and parameters
if score > dt_best_score:
dt_best_score = score
dt_best_parameter = md
# Rebuild a model on the combined training and validation set
SelectedDTModel = DecisionTreeClassifier(max_depth = dt_best_parameter).fit(X_train_scaled, Y_train )
# Model Test
dt_test_score = SelectedDTModel.score(X_test_scaled, Y_test)
# Predicted Output of Model
PredictedOutput = SelectedDTModel.predict(X_test_scaled)
# Extracting sensitivity & specificity from ROC curve to measure Performance of the model
dt_fpr, dt_tpr, dt_thresholds = roc_curve(Y_test, PredictedOutput, pos_label=1)
# Using AUC of ROC to validate models based on a single score
dt_test_auc = auc(dt_fpr, dt_tpr)
# Output Printing Scores
print("Best accuracy on validation set is:", dt_best_score)
print("Best parameter for the maximum depth is: ", dt_best_parameter)
print("Test accuracy with best parameter is ", dt_test_score)
print("Test AUC with the best parameter is ", dt_test_auc)
# Appending results to performance list
m = 'Decision Tree'
performance.append([m, dt_test_score, dt_test_auc, dt_fpr, dt_tpr, dt_thresholds])
Features Importance¶
In [29]:
print("Feature importance: ")
features = np.array([X.columns.values.tolist(), list(SelectedDTModel.feature_importances_)]).T
for i in features:
print(f'{i[0]} : {i[1]}')
Random Forest Classifier Model¶
In [30]:
rf_best_score = 0
rf_kfolds = 5
for M in range(2, 15, 2): # combines M trees
for d in range(1, 9): # maximum number of features considered at each split
for m in range(1, 9): # maximum depth of the tree
# train the model
# n_jobs(4) is the number of parallel computing
forestModel = RandomForestClassifier(n_estimators = M, max_features = d, n_jobs = 4,max_depth = m, random_state = 0)
# perform cross-validation
scores = cross_val_score(forestModel, X_train_scaled, Y_train, cv = rf_kfolds, scoring = 'accuracy')
# compute mean cross-validation accuracy
score = np.mean(scores)
# if we got a better score, store the score and parameters
if score > rf_best_score:
rf_best_score = score
rf_best_M = M
rf_best_d = d
rf_best_m = m
# Rebuild a model on the combined training and validation set
SelectedRFModel = RandomForestClassifier(n_estimators=rf_best_M, max_features=rf_best_d,max_depth=rf_best_m, random_state=0).fit(X_train_scaled, Y_train )
# Model Test
rf_test_score = SelectedRFModel.score(X_test_scaled, Y_test)
# Predicted Output of Model
PredictedOutput = SelectedRFModel.predict(X_test_scaled)
# Extracting sensitivity & specificity from ROC curve to measure Performance of the model
rf_fpr, rf_tpr, rf_thresholds = roc_curve(Y_test, PredictedOutput, pos_label=1)
# Using AUC of ROC to validate models based on a single score
rf_test_auc = auc(rf_fpr, rf_tpr)
# Output Printing Scores
print("Best accuracy on validation set is:", rf_best_score)
print("Best parameters of M, d, m are: ", rf_best_M, rf_best_d, rf_best_m)
print("Test accuracy with the best parameters is", rf_test_score)
print("Test AUC with the best parameters is:", rf_test_auc)
# Appending results to performance list
m = 'Random Forest'
performance.append([m, rf_test_score, rf_test_auc, rf_fpr, rf_tpr, rf_thresholds])
AdaBoost Classifier Model¶
In [31]:
ada_best_score = 0
ada_kfolds = 5
for M in range(2, 15, 2): # combines M trees
for lr in [0.0001, 0.001, 0.01, 0.1, 1,2,3]:
# train the model
boostModel = AdaBoostClassifier(n_estimators=M, learning_rate=lr, random_state=0)
# perform cross-validation
scores = cross_val_score(boostModel, X_train_scaled, Y_train, cv = ada_kfolds, scoring = 'accuracy')
# compute mean cross-validation accuracy
score = np.mean(scores)
# if we got a better score, store the score and parameters
if score > ada_best_score:
ada_best_score = score
ada_best_M = M
ada_best_lr = lr
# Rebuild a model on the combined training and validation set
SelectedBoostModel = AdaBoostClassifier(n_estimators=ada_best_M, learning_rate=ada_best_lr, random_state=0).fit(X_train_scaled, Y_train )
# Model Test
ada_test_score = SelectedBoostModel.score(X_test_scaled, Y_test)
# Predicted Output of Model
PredictedOutput = SelectedBoostModel.predict(X_test_scaled)
# Extracting sensitivity & specificity from ROC curve to measure Performance of the model
ada_fpr, ada_tpr, ada_thresholds = roc_curve(Y_test, PredictedOutput, pos_label=1)
# Using AUC of ROC to validate models based on a single score
ada_test_auc = auc(ada_fpr, ada_tpr)
# Output Printing Scores
print("Best accuracy on validation set is:", ada_best_score)
print("Best parameter of M is: ", ada_best_M)
print("best parameter of LR is: ", ada_best_lr)
print("Test accuracy with the best parameter is", ada_test_score)
print("Test AUC with the best parameters is:", ada_test_auc)
# Appending results to performance list
m = 'AdaBoost'
performance.append([m, ada_test_score, ada_test_auc, ada_fpr, ada_tpr, ada_thresholds])
XGBoost Classifier Model¶
In [32]:
xgb_best_score = 0
xgb_kfolds = 5
for n in [2,4,6,8,10]: #iterate over the values we need to try for the parameter n_estimators
for lr in [1.1,1.2,1.22,1.23,1.3]: #iterate over the values we need to try for the learning rate parameter
for depth in [2,4,6,8,10]: # iterate over the values we need to try for the depth parameter
XGB = xgb.XGBClassifier(objective = 'binary:logistic', max_depth=depth, n_estimators=n, learning_rate = lr) #define the model
# perform cross-validation
scores = cross_val_score(XGB, X_train_scaled, Y_train, cv = xgb_kfolds, scoring='accuracy')
# the training set will be split internally into training and cross validation
# compute mean cross-validation accuracy
score = np.mean(scores)
# if we got a better score, store the score and parameters
if score > xgb_best_score:
xgb_best_score = score #store the score
xgb_best_md = depth #store the parameter maximum depth
xgb_best_ne = n #store the parameter n_estimators
xgb_best_lr = lr #store the parameter learning rate
# rebuild a model with best parameters to get score
XGB_selected = xgb.XGBClassifier(objective = 'binary:logistic',max_depth=xgb_best_md, n_estimators=xgb_best_ne, learning_rate = xgb_best_lr).fit(X_train_scaled, Y_train)
# Model Test
xgb_test_score = XGB_selected.score(X_test_scaled, Y_test)
# Predicted Output of Model
PredictedOutput = XGB_selected.predict(X_test_scaled)
# Extracting sensitivity & specificity from ROC curve to measure Performance of the model
xgb_fpr, xgb_tpr, xgb_thresholds = roc_curve(Y_test, PredictedOutput, pos_label=1)
# Using AUC of ROC to validate models based on a single score
xgb_test_auc = auc(xgb_fpr, xgb_tpr)
# Output Printing Scores
print("Best accuracy on cross validation set is:", xgb_best_score)
print("Best parameter for maximum depth is: ", xgb_best_md)
print("Best parameter for n_estimators is: ", xgb_best_ne)
print("Best parameter for learning rate is: ", xgb_best_lr)
print("Test accuracy with the best parameters is", xgb_test_score)
print("Test AUC with the best parameters is:", xgb_test_auc)
# Appending results to performance list
m = 'XGB'
performance.append([m, xgb_test_score, xgb_test_auc, xgb_fpr, xgb_tpr, xgb_thresholds])
Ensemble Voting Classifier Model¶
In [33]:
from sklearn.ensemble import VotingClassifier
# Logistic regression model
clf1 = LogisticRegression(C=lr_best_parameters).fit(X_train_scaled, Y_train)
# SVC Model
clf2 = SVC(C=svm_best_parameter_c, gamma=svm_best_parameter_gamma, kernel=svm_best_parameter_k,probability=True)
# DecisionTreeClassifier model
clf3 = DecisionTreeClassifier(max_depth=dt_best_parameter)
# Random Forest Classifier Model
clf4 = RandomForestClassifier(n_estimators=rf_best_M, max_features=rf_best_d,max_depth=rf_best_m, random_state=42)
# AdaBoostClassifier Model
clf5 = AdaBoostClassifier(n_estimators=ada_best_M, learning_rate=ada_best_lr, random_state=42)
# XGBoost Classifier Model
clf6 = xgb.XGBClassifier(objective = 'binary:logistic',max_depth=xgb_best_md, n_estimators=xgb_best_ne ,learning_rate = xgb_best_lr)
# Defining VotingClassifier
eclf1 = VotingClassifier(estimators=[ ('LogisticRegression', clf1), ('SVC', clf2),('DecisionTree',clf3),('Random Forest', clf4),('ADABoost',clf5),('XGBoost',clf6)], voting='hard')
# Fitting VotingClassifier
eclf1 = eclf1.fit(X_train_scaled, Y_train)
# Model Test
eclf_test_score = eclf1.score(X_test_scaled, Y_test)
# Predicted Output of Model
PredictedOutput = eclf1.predict(X_test_scaled)
# Extracting sensitivity & specificity from ROC curve to measure Performance of the model
eclf_fpr, eclf_tpr, eclf_thresholds = roc_curve(Y_test, PredictedOutput, pos_label=1)
# Using AUC of ROC to validate models based on a single score
eclf_test_auc = auc(eclf_fpr, eclf_tpr)
# Output Printing Scores
print("Test accuracy with the best parameters is", eclf_test_score)
print("Test AUC with the best parameters is:", eclf_test_auc)
# Appending results to performance list
m = 'ECLF'
performance.append([m, eclf_test_score, eclf_test_auc, eclf_fpr, eclf_tpr, eclf_thresholds])
Model Results for Top performing models¶
In [34]:
result = pd.DataFrame(performance, columns=['Model', 'Accuracy', 'AUC', 'FPR', 'TPR', 'TH'])
df = result[['Model', 'Accuracy', 'AUC']]
results = df.sort_values('Accuracy',ascending = False)
display(results.style.background_gradient(cmap='Reds',subset=["Accuracy"]).background_gradient(cmap='Greens',subset=["AUC"]))
Conclusion¶
- Best Model Accuracy is SVM Classifier at ~94% accuracy.
- Best AUC is SVM Classifier at 86% area.
- Ensemble Voting is at ~92% accuracy and 81% area.
- Most contributing features are PPE at 35% followed by MDVP:Fo(Hz) at 28% and HNR at 13%.
- Machine Learning Findings align with EDA.